Learning to utilize internal protein 3D nanoenvironment descriptors in predicting CRISPR–Cas9 off-target activity

Despite advances in determining the factors influencing cleavage activity of a CRISPR-Cas9 single guide RNA (sgRNA) at an (off-)target DNA sequence, a comprehensive assessment of pertinent physico-chemical/structural descriptors is missing. In particular, studies have not yet directly exploited the information-rich internal protein 3D nanoenvironment of the sgRNA-(off-)target strand DNA pair, which we obtain by harvesting 634 980 residue-level features for CRISPR-Cas9 complexes. As a proof-of-concept study, we simulated the internal protein 3D nanoenvironment for all experimentally available single-base protospacer-adjacent motif-distal mutations for a given sgRNA-target strand pair. By determining the most relevant residue-level features for CRISPR-Cas9 off-target cleavage activity, we developed STING_CRISPR, a machine learning model delivering accurate predictive performance of off-target cleavage activity for the type of single-base mutations considered in this study. By interpreting STING_CRISPR, we identified four important Cas9 residue spatial hotspots and associated structural/physico-chemical descriptor classes influencing CRISPR-Cas9 (off-)target cleavage activity for the sgRNA-target strand pairs covered in this study.